基于少样本学习算法的结直肠粘膜下肿瘤和息肉内镜图像分类系统  

Development of a few-shot learning based model for the classification of colorectal submucosal tumors and polyps on endoscopic images

作　　者：伍亚辉 朱世祺[1] 吴宇东 张儒发 朱锦舟 WU Yahui;ZHU Shiqi;WU Yudong;ZHANG Rufa;ZHU Jinzhou(Department of Gastroenterology,the First Affiliated Hospital of Soochow University,Suzhou 215006,China;Department of Pediatric Internal Medicine,Shanghai East Hospital,Tongji University,Shanghai 200120,China;Department of Internal Medicine,Suzhou Xiangcheng District Yangcheng Lake People's Hospital,Suzhou 215138,China;Department of Gastroenterology,Changshu Hospital Affiliated to Soochow University,Suzhou 215500,China)

机构地区：[1]苏州大学附属第一医院消化内科,江苏苏州215006 [2]同济大学附属东方医院儿内科,上海200120 [3]苏州市相城区阳澄湖人民医院内科,江苏苏州215138 [4]苏州大学附属常熟医院消化内科,江苏苏州215500

出　　处：《中国医学物理学杂志》2024年第7期897-904,共8页Chinese Journal of Medical Physics

基　　金：国家自然科学基金(82000540);肝脾外科教育部重点实验室开放基金(GPKF202304);苏州市科教兴卫项目(KJXW2019001);苏州大学医学部学生课外科研项目(2021YXBKWKY050)。

摘　　要：目的:为解决难以收集足够结直肠粘膜下肿瘤样本训练深度学习模型的问题,基于少样本学习算法构建了粘膜下肿瘤和息肉内镜图像分类模型。方法:收集多中心来源的结直肠粘膜下肿瘤内镜图像共172张,包括结直肠脂肪瘤(CRLs)、神经内分泌肿瘤(NETs)、锯齿状病变及息肉、传统腺瘤各43张。基于这些内镜图像构建支持集和查询集,在ImageNet和食管内镜图像上二次预训练的ResNet50提取图像特征,计算欧氏距离,使用K近邻算法进行分类。与原始模型和低、高年资内镜医师进行对比,评估少样本学习模型的分类性能。结果:提出的少样本学习模型分类准确率、宏曲线下面积和MacroF1值分别为0.831、0.925和0.831,诊断CRLs的准确率和F1值分别为0.925和0.850,诊断NETs的准确率和F1值分别为0.906和0.805。同时,该模型具有较好的分类一致性(Kappa=0.775)和可解释性。结论:构建的少样本学习模型在区分CRLs、NETs、锯齿状病变及息肉、传统腺瘤内镜图像上表现出优异性能,可用于辅助内镜下识别结直肠粘膜下肿瘤。Objective To address the difficulty in collecting sufficient endoscopic images of colorectal submucosal tumors for traditional deep learning model training,a few-shot learning based model(FSL model)is proposed for classifying colorectal submucosal tumors and polyps on endoscopic images.Methods A total of 172 endoscopic images of colorectal submucosal tumors were collected from different centers,including 43 each of colorectal lipomas(CRLs),neuroendocrine tumors(NETs),serrated lesions and polyps(SLPs),and traditional adenomas.A support set and a query set were constructed using these endoscopic images.ResNet50 which was pre-trained on ImageNet and esophageal endoscopic images was used to extract image features.Subsequently,K-nearest neighbors algorithm was used for classification based on the calculated Euclidean distance.The classification performance of FSL model was evaluated through the comparison with the original model and endoscopists.Results FSL model had a 4-class classification accuracy of 0.831,Macro AUC of 0.925,Macro F1-score of 0.831;moreover,the proposed model achieved diagnostic accuracies of 0.925 and 0.906 for CRLs and NETs,with F1 score of 0.850 and 0.805.Additionally,the proposed model exhibited high classification consistency(Kappa=0.775)and interpretability.Conclusion The established FSL model performs well in distinguishing CRLs,NETs,SLPs and traditional adenomas on endoscopic images,indicating its potential utility in assisting the identification of colorectal submucosal tumors under endoscopy.

关 键 词：少样本学习 结直肠粘膜下肿瘤 结直肠息肉 消化内镜图像 深度学习 

分 类 号：R318[医药卫生—生物医学工程] R574[医药卫生—基础医学]